--- Log opened Wed Jul 30 00:00:55 2025
00:07 -!- TMM [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.]
00:07 -!- TMM [hp@amanda.tmm.cx] has joined #hplusroadmap
01:58 -!- darsie [~darsie@84-113-82-174.cable.dynamic.surfer.at] has quit [Quit: Avoid fossil fuels and animal products. Have no/fewer children. Protest, elect sane politicians. Invest ecologically.]
01:59 -!- darsie [~darsie@84-113-82-174.cable.dynamic.surfer.at] has joined #hplusroadmap
05:38 -!- Hooloovoo [~Hooloovoo@hax0rbana.org] has quit [Ping timeout: 260 seconds]
05:40 -!- Hooloovoo [~Hooloovoo@hax0rbana.org] has joined #hplusroadmap
06:54 -!- justanotheruser [~justanoth@gateway/tor-sasl/justanotheruser] has joined #hplusroadmap
09:14 < kanzure> "Keybase, but with federation, SSO and YubiKey support, and it's open-source" https://blog.foks.pub/posts/introducing/ https://foks.pub/
09:27 < kanzure> it's a key-value store bolted on top of git? why?
09:28 < kanzure> git isn't a bad choice but what's the rest of this for?
09:29 < kanzure> is it simply a wrapper around PGP or GPG for people who don't want to be seen directly using GPG?
09:31 < kanzure> and if so why is it not advertised as such.
10:03 -!- darsie [~darsie@84-113-82-174.cable.dynamic.surfer.at] has quit [Quit: Avoid fossil fuels and animal products. Have no/fewer children. Protest, elect sane politicians. Invest ecologically.]
10:04 -!- darsie [~darsie@84-113-82-174.cable.dynamic.surfer.at] has joined #hplusroadmap
11:32 < kanzure> no comments on the log summaries?
11:55 < superkuh> It's fun nostalgia at the least.
11:56 < kanzure> looking for feedback so i can instruct a minion on what to do differently. and/or build other things, such as web search over the logs. or something.
12:00 < RangerMauve> Maybe do some link agregation and build up the wiki with knowledg extracted from the chats. Maybe a separate wiki to start to compare effectiveness with human curation
12:06 < kanzure> ah, separate wiki. okay.
12:07 < kanzure> prompt would be something like "find at least one appropriate wiki page for each conversation and propose an update to each matching page based on any new important information from that conversation in the log"?
12:08 < kanzure> blockquoting chat excerpts on a wiki and anchor-linking back to htmlized logs would be cool.
12:18 < RangerMauve> I've had some ideas for multi shot wiki curation / search but I'm not sure if it applies to large models that can auto-magic everything.
12:49 < jrayhawk> if claude puts enough wrong things in the wiki, will we feel compelled to replace those wrong things with right things?
12:53 -!- darsie [~darsie@84-113-82-174.cable.dynamic.surfer.at] has quit [Quit: Avoid fossil fuels and animal products. Have no/fewer children. Protest, elect sane politicians. Invest ecologically.]
12:54 -!- darsie [~darsie@84-113-82-174.cable.dynamic.surfer.at] has joined #hplusroadmap
13:07 < kanzure> jrayhawk: well RangerMauve's proposal was a separate wiki
14:05 < RangerMauve> When talking about the approach with a friend whose whole thing is knowledgebase creation/curation they were skeptical that the LLM would give good results. Might be fine though
14:08 -!- Gooberpatrol66 [~Gooberpat@user/gooberpatrol66] has quit [Quit: Konversation terminated!]
14:17 -!- Gooberpatrol66 [~Gooberpat@user/gooberpatrol66] has joined #hplusroadmap
14:28 < hprmbridge> kanzure> got another strategy?
14:28 < hprmbridge> norkatron> the biggest issue i've run into is sometimes they skip the middle of a document you are trying to summarize , besides that summarization is one of the big wins of LLMs imo ..
14:28 < hprmbridge> norkatron>
14:28 < hprmbridge> norkatron> been wanting to try out for a while various methods of chunking up text to avoid this problem , would be interesting to find metrics on if there are specific things an llm will gloss over vs not gloss over or how much txt length and how much txt length they can handle
14:29 < hprmbridge> kanzure> one way might be comparing an entropy metric of the text based off of word frequency against the summary entropy or frequency.
14:30 < L29Ah> summarization is the only LLM use case i productionalized for myself
14:31 < L29Ah> LLM doctor will likely go second
14:35 < L29Ah> although i'm sure if you have a lot of compute you can use LLM for hypothesis generation and prioritization productively
14:37 < L29Ah> 20:56:21]<kanzure> looking for feedback so i can instruct a minion on what to do differently. and/or build other things, such as web search over the logs. or something.
14:37 < L29Ah> make the minion find and list the most important things over week/month/year
14:37 < L29Ah> *LLMinion
14:38 < RangerMauve> Regarding summarization / wiki building I've been thinking of a divide and conquer method. Split the data into chunks (chapters, pages, sections) and have it extract bits of relevant knowledge from there. Then build that up into a larger chunk to think over.
14:39 < RangerMauve> For wiki I've been thinking of making it traverse the wiki by searching for pages and listing headings, then pulling specific sections out and patching onto those or adding new sections
14:40 < RangerMauve> Dividing also opens the door to paralell processing of data extraction
14:48 < hprmbridge> norkatron> could definitely be a step towards testing how well an llm is in fact summarizing hmm
15:02 -!- darsie [~darsie@84-113-82-174.cable.dynamic.surfer.at] has quit [Quit: Avoid fossil fuels and animal products. Have no/fewer children. Protest, elect sane politicians. Invest ecologically.]
15:03 -!- darsie [~darsie@84-113-82-174.cable.dynamic.surfer.at] has joined #hplusroadmap
15:16 < hprmbridge> kanzure> https://herasight.com/
15:17 < hprmbridge> kanzure> and their announce thread https://x.com/AlexTISYoung/status/1950575617294180510
15:31 -!- TMM [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.]
15:32 -!- TMM [hp@amanda.tmm.cx] has joined #hplusroadmap
16:52 < hprmbridge> Eli> What do you think about their iq claims?
16:55 < hprmbridge> .monokhrome> How can they be sure about their predictors if the people haven't been born and gotten old enough to develop the diseases?
16:56 < hprmbridge> kanzure> you might be asking "does GWAS and polygenics have any epistemic or epistemological statistical validity?"
16:56 < L29Ah> it is easy if you plan to sell the company or go bankrupt in the next ten years
16:56 < hprmbridge> .monokhrome> Snart
16:57 < hprmbridge> .monokhrome> *Smart
16:58 < hprmbridge> kanzure> their claims seem to be statistically reasonable to me.
16:58 < hprmbridge> kanzure> but personally I'm more interested in genetic engineering
16:59 < hprmbridge> kanzure> why select from what's already there? why not just... do the thing you actually want.
17:00 < L29Ah> you can earn some for lobbying that while you're at your selection thing
17:00 < hprmbridge> .monokhrome> The problem is the long development cycle. It's going to take 20 years to find out if your gene modifications boosted adult iq or gave the person schizophrenia, then you have to go back to the drawing board and tweak it and wait another 20 years to see if you've fixed it. And any patents you took out on the genes will have expired.
17:01 < L29Ah> and then be the first one to do engineering with most of the processes already in place and perfected
17:02 < L29Ah> or go broke when you figure nobody cares even about selection, so you won't have to invest into lobbying for engineering
17:02 < hprmbridge> .monokhrome> And there's an even longer cycle for life extension traits
17:21 -!- stipa_ [~stipa@user/stipa] has joined #hplusroadmap
17:23 -!- stipa [~stipa@user/stipa] has quit [Ping timeout: 252 seconds]
17:23 -!- stipa_ is now known as stipa
17:23 < hprmbridge> kanzure> it's not going to take 20 years.
17:24 < hprmbridge> kanzure> IQ testing begins at 2 or 3 years old. children as young as 4 years old have been sent to work and do labor. 20 years is ridiculous.
17:25 < hprmbridge> kanzure> but separately, we can use our knowledge to not have to wait that long either
17:57 < hprmbridge> kanzure> have you some open weights for great win: Deepseek (Dec), Gemma 3 (Mar), OLMO 2 (Mar), Qwen3 (Apr), Falcon H1 (May), Hunyuan A13B (Jun), Kimi K2 (Ju), Kimi Dev 72B (Ju), Qwen3-Coder (Jul) or GLM 4.5 (Jul)
18:02 -!- darsie [~darsie@84-113-82-174.cable.dynamic.surfer.at] has quit [Quit: Avoid fossil fuels and animal products. Have no/fewer children. Protest, elect sane politicians. Invest ecologically.]
18:03 -!- darsie [~darsie@84-113-82-174.cable.dynamic.surfer.at] has joined #hplusroadmap
18:41 < fenn> i like the auto wiki idea. i think i talked about it before, possibly even in here
18:43 < fenn> fenn> "take this day's chat logs, find a relevant wiki article or start a new one if there isn't one, and integrate the conclusions of the day's discussion in the appropriate place in the wiki page"
18:45 < fenn> it would probably be better to make an embedding with a big model like Qwen3-Embedding-8B, lay it out with t-SNE, manually identify clusters and create a skeleton of categories from them, and THEN set the LLM loose on it
18:45 < fenn> that way you have the entirety of the chat logs in the category making process, which is impossible for a limited context window to do
18:47 < hprmbridge> kanzure> how about just good semantic vector search
18:48 < fenn> that's literally a search engine
18:48 < fenn> is that what you want?
18:49 < fenn> once you have the embeddings for the entire history it "should" be easy to do
18:55 < fenn> Muaddib: say we collect a bunch of naturally occurring alleles that all synergistically add together, and put them in the same organism with DNA editing techniques. is this natural or modified? can the resulting organism or its selection of genes be patented?
18:55 < Muaddib> refusal, rip
18:55 < fenn> lame
19:02 < fenn> kanzure: re log summaries, i don't particularly care for the passive tense "A discussion about..." "There was mention of..." "...was shared" it just adds unnecessary verbiage 
19:03 < fenn> gwillen did a bunch of work on getting the best prompt for claude, what happened to that code / prompts?
19:04 < fenn> it should be about the topics, not the fact that participants talked about the topics
19:05 < fenn> also it needs to be normalized to chat volume for that day, or something
19:07 < fenn> you should filter out joins/parts before processing since 99.99% of the time that's not relevant to the discussion
19:09 < fenn> i don't understand what went wrong here: https://gnusha.org/logs/summaries/2025-07-28/2017-02-10.md vs https://gnusha.org/logs/2017-02-10.log
19:12 < fenn> maybe we want to do something like named entity extraction to create wiki pages
19:14 < hprmbridge> kanzure> please speculate in the form of properly stated LLM prompts for a junior developer to execute on
19:14 < hprmbridge> kanzure> there's a real junior, it's not just another llm
19:19 < fenn> this is something that will have to be iterated on and looked at, but, something like:
19:23 < fenn> from the above chat log, list a markdown table consisting of short phrases suitable for a wiki article name, which uniquely identify each subject, person being discussed, theory, concept, place, program, technology, gene, product, service, law, meme, and so on
19:23 < hprmbridge> kanzure> https://diyhpl.us/wiki/chatsummaries/
19:24 < fenn> then we go back through and ask it "is there anything important being discussed which is not in the following list of topics?" and feed it the previous generated list of phrases
19:24 < fenn> for all of my LLM stuff i try to keep the long document at the start of the context window and put the questions at the end, so the prefix is still in cache
19:25 < fenn> i don't know if the big providers bill this correctly or not
19:28 < fenn> only two of those documents are summaries
19:29 < fenn> i much prefer the 2014-05-15 one
19:31 < hprmbridge> kanzure> that was me
19:31 < hprmbridge> kanzure> trying to write a good example
19:33 < hprmbridge> kanzure> https://diyhpl.us/wiki/projects/human-like-cognitive-abilities/ man I can't believe I got the parrot part right. and the mentally challenged part was right too.
19:37 < hprmbridge> kanzure> i should write a page like that for human IQ enhancement and start the hyperstition going in that direction
19:39 < fenn> "if this person is added to the project, will the probability of success increase by at least 1,000x?" you invented the 1000x developer concept too
19:40 < hprmbridge> kanzure> yep, that's why zuck is making $1 billion dollar offers
19:40 < fenn> yeesh
19:40 < fenn> it's great the engineers are finally being valued, but, uh... this doesn't really work for society
19:41 < fenn> i think our wiki knitting bot will have to actually read or at least skim the links to make sense of what's being discussed
19:42 < hprmbridge> kanzure> ok maybe each link can be crawled and its content interlaced right there in the log
19:42 < fenn> that implies some program that scrapes the links and converts them to text and gets rid of junk before adding it to the context
19:43 < fenn> i'm not sure about summarizing the links because sometimes people reference specific phrases in the linked document word for word
19:43 < fenn> i know i do anyway
19:45 < fenn> there's a lot of links and the total quantity of linked text is much greater than the discussion text, and this would be quite distracting. maybe it'd be better to extract a snippet around any verbatim quotes and include those along with a summary
19:46 < fenn> and then you can do the thing in the html version of the logs where it shows you some preview summary on mouseover of the link
19:46 < hprmbridge> kanzure> for manuscripts instead of a summary you can extract its abstract
19:47 < fenn> i want to believe
20:00 < fenn> i vaguely remember some idea i had where each synthetic data point or line of autogenerated text is shown in one color, then it gets hue shifted the more attestations of its correctness or incorrectness are added to it
20:01 < fenn> bots and humans can have different colors or styles like say bots only change the color of the text and humans change the color of the underline
20:03 < hprmbridge> kanzure> trillions will die for the king claude annotated version.
20:03 < fenn> the qwenarians will see too it! infidel!
20:07 < fenn> https://fennetic.net/irc/homeworld%20deserts%20of%20kharak%20gaalsien%20leader%206bddc29b06c34e77b16352f8f750a910aa88ae6f.jpg
20:08 < hprmbridge> kanzure> https://x.com/kanzure/status/1950733468431769899 "add more grey matter" is something we know how to do. I think the trouble is more that people want to skip the experimentation and testing stage. They want a sure thing. Who is investing in early stage, high risk intelligence enhancement interventions? Not big pharma..
20:09 < fenn> can you do superfrogs without a lot of red tape? since nobody cares about slimy green people
20:10 < hprmbridge> kanzure> yes
20:10 < hprmbridge> kanzure> funny, i was just discussing that with someone, seeing if he wants to run the frog farm
20:11 < fenn> is the goal to have one giant brain frog or a network or frogs
20:11 < hprmbridge> kanzure> actually a year ago i had an investor offer to fund a frog breeding intelligence program with a small first angel investment check....
20:11 -!- TMM [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.]
20:11 -!- TMM [hp@amanda.tmm.cx] has joined #hplusroadmap
20:12 < hprmbridge> kanzure> a population of smart uplifted animals that are individually agentic intelligences
20:12 < fenn> i bet humans have a lot of adaptations to handle the large amount of DHA and cholesterol needed for large brains
20:12 < fenn> so a bunch of giant brain frogs?
20:13 < hprmbridge> kanzure> even the dumb version of this that takes 20 years should be done. shortcuts like engineering, genetics, etc are bonuses.
20:13 < hprmbridge> kanzure> yeah sure
20:14 < fenn> should we let the frogs know that they will be destroyed if the FDA finds out about their existence?
20:15 < hprmbridge> kanzure> frog legs are legal food
20:15 < fenn> but the frog is a drug, for the frog
20:16  * fenn scratches his head
20:16 < fenn> i'm sure it makes sense to somebody!
20:16 -!- darsie [~darsie@84-113-82-174.cable.dynamic.surfer.at] has quit [Ping timeout: 248 seconds]
20:17 < fenn> this guy unironically says "there are decreasing returns to IQ"
20:18 < fenn> no, what you meant is the product of IQ and other stuff like perseverance is what's valuable
20:21 < fenn> so Stephen Jay Gould wrote this book about how victorian scientists measuring skull volume was a case of pathological science and they made up all the data to fit their preconceived notions, "The Mismeasure of Man"
20:21 < fenn> but then it turned out that Stephen Jay Gould was the one engaging in scientific fraud and ... i forget the rest
20:21 < fenn> so now i'm unclear on the relationship between brain volume and IQ
20:23 < fenn> just looking at some people in high falutin academic positions and you can see they have huge heads, but this may be about knowledge capacity rather than IQ
20:25 < fenn> it has to be that IQ is a product of multiple other factors, because we're made out of biology and it has a lot of moving parts
20:30 < fenn> n=1349 siblings, measure head circumference and IQ, β ± SE = 0.172 ± 0.031, ρ=0.19, p-value=3e-8
20:30 < fenn> so a weak effect
20:31 < fenn> claude says it explains 3% to 4% of the variation (0.19^2)
20:35 < hprmbridge> Eli> The FDA needs to recognize being stupid as a disease. I'll be first in line for the treatment.
20:36 < fenn> The FDA needs to recognize being stupid
22:31 < hprmbridge> kanzure> who cares if head size presently does or does not correlate. what? literally doesn't matter. if you had a +IQ intervention that increased head size, would you have to abandon it because it isn't supported by unrelated data?
22:35 < fenn> engineering is about tradeoffs. there are costs to having a large head, and if we could find more powerful interventions without those costs, it would be a better strategy. pareto frontier blah blah blah
22:36 < fenn> it would be better to start with birds :\
22:42 < fenn> https://paste.mathizen.net/uploads/Oux4JM5pxH.jpeg
22:56 -!- Gooberpatrol66 [~Gooberpat@user/gooberpatrol66] has quit [Quit: Konversation terminated!]
23:43 -!- _flood [~flooded@193.37.254.189] has quit [Remote host closed the connection]
23:44 -!- _flood [~flooded@193.37.254.189] has joined #hplusroadmap
23:44 -!- _flood [~flooded@193.37.254.189] has quit [Read error: Connection reset by peer]
23:45 -!- _flood [~flooded@193.37.254.189] has joined #hplusroadmap
--- Log closed Thu Jul 31 00:00:56 2025